Table Of Content

~ld AD A254m 802wi n~ldwi~TeIwzmiu NEURAL NETWORKS AND NON-DESTRUCTIVE TEST/EVALUATION METHODS by JEFFREY DEAN DRAPER A Scholarly Paper submitted to ASSISTANT PROFESSOR IAN FLOOD of THE UNIVERSITY OF MARYLAND, COLLEGE PARK for Partial fulfillment of the Requirements for the Degree of Master of Science in Civil Engineering - For 1992 -For , :-. - -, a ,t --. D'TIC QUALrrTy rNlfpCsT'R 8 -" ;- -- 1 1 K' 'A Neural Networks & Non-Detruive Test/Evaluation Methods TABLE OF CONTENTS ABSTRACT ............................................. Page 1 I. INTRODUCTION ..................................... Page 1 II. LITERATORE REVIEW ............................... Page 4 III. AIMS AND OBJECTIVES OF THE RESEARCH ........... Page 10 IV. RESEARCH METHODOLOGY ........................ Page 11 V. ARTIFICIAL NEURAL NETWOP.K BASICS .............. Page 13 VI. APPLICATION DEVELOPMENT OF THE ANN FOR VISUAL IMAGE ENHANCER AND FILTER PROBLEM ........... Page 19 VII. CONCLUSIONS & RECOMMENDATIONS ............... Page 30 VIII. ACKNOWLEDGEMENTS ............................ Page 32 APPENDIX I. SPECIFIC TESTS USED TO EVALUATE CONCRETE ............................ Page 33 APPENDIX II. PROGRAM PATTERN3 FOR THE DEVELOPMENT OF TRAINING PATTERNS FOR A 3 X 3 SAMPLING WINDOW ......... Page 34 APPENDIX III. PROGRAM PATTERNS FOR THE DEVELOPMENT OF TRAINING PATTERNS FOR A 9 X 9 SAMPLING WINDOW ......... Page 40 APPENDIX IV. THE TEMPLATE FOR THE PARAMETERS OF THE 3 X 3 SAMPLING WINDOW AS DEVELOPED BY THE PROGRAM BINARYHAM ........................... Page 46 APPENDIX V. THE TEMPLATE FOR THE PARAMETERS OF THE 9 X 9 SAMPLING WINDOW AS DEVELOPED BY THE PROGRAM BINARYHAM .......................... Page 47 Draper -April 1992 Page i Neural Nenwodak & Non-Destuctive Test/Evaluaton Methods TABLE OF CONTENTS (Continued) APPENDIX VI. NOTATIN ............................. Page 48 APPENDIX VII. REFERENCES...........................Page 49 Draper - April 1992 Page ii Neural Networks & Non-Destructiv Test/Evaluation Methods LIST OF FIGURES Figure 1: CONCRETE NDTE METHODS ..................... Page 9 Figure 2: SCHEDULE OF WORK ........................... Page 12 Figure 3: THREE-LAYER FEEDFORWARD NEURAL NETWORK Page 15 Figure 4: COMMON NEURON TRANSFER FUNCTIONS ........ Page 17 Figure 5: NEURAL NETWORK TRAINING METHODS ......... Page 18 Figure 6: MAJOR PHASES OF ANN APPLICATION DEVELOPMENT ................................. Page 19 Figure 7: SAMPLE TRAINING PATTERNS FOR THE 3 X 3 SAMPLING WINDOW ............................. Page 24 Figure 8: EXAMPLE TRAINING PATTERNS FOR THE 9 X 9 SAMPLING WINDOW ............................. Page 25 Figure 9: ORIGINAL VS. ANN FILTERED/ENHANCED IMAGES FOR THE 3 X 3 SAMPLING WINDOW ............... Page 28 Figure 10: ORIGINAL VS. ANN FILTERED/ENHANDED IMAGE FOR 9 X 9 SAMPLING WINDOW .................... Page 29 Draper -April 1992 Page iii Neural Networks & Non-Destuctve Test/Evaluation Methods ABSTRACT: With today's reports of deterioratingh ighways and infrastructurea s well as increased litigation arisingf rom structuralf ailures and the construction process, there is an increasingd esire to employ non-destructive testing and evaluation (NDTE) methods for analyzing structuralc oncrete members as well as other construction materials in a noninvasive manner. A major part of NDTE techniques is defect characterization,w hich is a typical pattern classification problem. The current state of the art for solving this problem is the application of a human expert's knowledge and experience for interpreting NDTE data Artificial neural networks (ANNs) have shown a propensity for solving the pattern classificationp roblem in the areas of speech and vision recognition, as well as problems in system modeling and simulation. As a result of these successful ANN applications, this paper explores the possibility of using ANNs for the NDTE defect characterizationp roblem. Part of the solution of defect characterizatione ntails the capability to filter what would otherwise be considered noisy data. Therefore, an ANN architecturei v proposed and tested via computer simulation for the purpose of discerning between cracks and other surface defects found in photographs of defective reinforced concrete sections. Also, a basic introduction to ANNs is included along with a recommendation for continuing research. I. INTRODUCTION Reinforced concrete generally performs well as long as conditions for its installation and use fall within the parameters for which it was designed. However, there have been and always will be occasions involving severe construction conditions, construction mistakes, faulty design, unforeseen disasters such as fire and flood, and/or unanticipated loads placed on structural concrete. As a result of these aforementioned inauspicious circumstances, a reinforced concrete member will show signs of distress, i.e. cracking, dusting, scaling, spauling, etc. These signs of distress will require either one or some combination of the owner, designer, and/or constructor to investigate the reinforced member to determine its strength, anticipated longevity, and need for Draper - April 1992 Page 1 Neural Newodrs & Non-Dlsuaive Test/Evaluation Methods replacement. Indisputably, it is ideal with respect to time and money to investigate the structure without doing any damage to the member; for this reason, non-destructive test and evaluation (NDTE) methods have become popular and necessary means for analyzing the integrity of structural concrete. Like many other scientific techniques, NDTE heavily relies on some expert to collect, graph, and interpret data. Certainly, there will never be an engineering tool which will eliminate the need for experts and good judgement. However, automation of NDTE methods would improve the speed of analyses and likely increase the frequency with which these methods are used. At a minimum, NDTE automation would allow "non-experts" who become trained on automated NDTE systems to engage in initial data collection and defect classification part of the problem. NDTE data collection and interpretation is generally a problem in pattern classification, i.e. a true expert would almost instantly recognize that data from any given situation fits some particular problem and solution method which he has before seen. However, pattern classification lends itself poorly to traditional computing methods. Conventional computer pattern classification has involved feature extraction and clustering which more often than not requires the use of extensive prior information, such as the statistical distribution of vectors. Draper - April 1992 Page 2 Neural Networks & Non-Deasudve Test/Evaauion Methods In the case of NDTE data pattern classification, the computerization problem is compounded with the fact that the cause for and impact of defects on materials, like many other real world functions, is extremely complex to model, requires the consideration of many factors (independent variables), and is not completely understood. Enter the artificial neural network. An artificial neural network (ANN) is either a hardware or software system which attempts to imitate the neural structure and functioning of the biological brain (Sejnowski, Kock, and Churchland, 1988). The brain uses millions of elementary processors known as neurons which are interconnected by synapses and process sensory information (sight, sound, touch, smell, and taste), thus allowing us to perceive and ultimately react to our environment. Similar to its biological counterpart, an ANN is a massively parallel, interconnected network of simple processors which can receive and process many independent variables. The basic advantage of the ANN over other traditional serial computing techniques is the ability to take into account and process many independent variables much faster. In addition, ANNs have shown promise for successfully performing a variety of cognitive tasks, including statistical pattern classification. Practical applications of ANNs as pattern classifiers and the need for real-time response to real-world data have led to advances in automated speech recognition, vision Draper -April 1992 Page 3 Neul Ndetw & Non-Demucdve Test/Evaluation Methods recognition, robotics, and other various engineering and artificial intelligence applications. NDTE defect classification is similar to the aforementioned real-world cognition problems for which ANNs have already been shown to have promise, i.e. problems requiring the processing of many independent variables and the classification of the result. By combining ANNs with NDTE, a significant improvement is expected in the consistency, accuracy, and ease of classifying NDTE data, i.e., a fraction of a second to classify an x-ray image or surface photograph. II. LITERATURE REVIEW Despite the initial skepticism in their applications and abilities (Minsky and Papert, 1969), ANNs have been shown by contemporary research as capable of solving a variety of engineering problems. The list of ANN applications includes: * Classification of speech sounds (Lippman, 1987), * Recognition of incoming military targets (Roth, 1990), " Formation of text-to-phoneme rules (Sejnowski and Rosenberg, 1987), Draper - April 1992 Page 4 Neural Network & Non-Daftidve Test/Evahufaon Methods " Deduction of the secondary structure of a protein from its amino acid sequence (Qian and Sejnowski, 1988), * Discrimination between underwater sonar signals (Gorman and Sejnowski, 1988), * Recognition of handwriting (Weideman, Manry, and Yau, 1989), * Learning good moves for backgammon (Tesauro and Sejnowski, 1988), * Performance of nonlinear signal processing (Lippman and Beckman, 1989; Tamura and Waibel, 1988), " Prediction of the amount of energy needed to modify the thermal energy stored in a building mass (Garret, et al, 1991), * Controlling the threshing module of a combine harvester (Garret, et al, 1991), * Design of pump locations and rates of operation (Garret, et al, 1991), " Recognition of machining features from a CAD drawing (Garret, et al, 1991), With the successes of the back-propagation neural network classifier (Rumelhart, McClleland, et al, 1986) and other various ANN forms, the field of construction engineering and management has been also been targeted as an area rich with potential ANN applications (Mohan, 1990). Some proposed Draper - April 1992 Page 5